Counterfactual Diagnostics Separate Visual Sensitivity from Claim Persistence in LVLMs

A diagnostic tool reveals that 27.81% of correct Qwen2.5-VL-7B responses contain unsupported visual claims. The method distinguishes between Evidence-Function Sensitivity (prediction changes) and claim persistence (the tendency to maintain false claims) to improve credit assignment in multimodal RL.